Environment Variables

Every auxilia deployment is configured through environment variables. This page is the complete reference — start here to understand what to set, then follow the Cloud Run or VM / Docker host guide for the deployment-specific wiring.

A full example lives in .env.example at the repo root.

Required variables

These must be set for auxilia to start. Missing any of them will fail backend boot.

Variable	Notes
`DATABASE_URL`	`postgresql+psycopg://user:pass@host:5432/dbname` connection string
`REDIS_HOST` / `REDIS_PORT`	Reachable from the backend (OAuth tokens are cached here)
`JWT_SECRET_KEY`	32+ random characters — used to sign session cookies
`SALT`	Salt for AES-GCM encryption of stored MCP API keys
`FRONTEND_URL`	Public URL of the web app (used as OAuth redirect base)
`BACKEND_URL`	Internal URL the Next.js proxy uses to reach the backend (web only)
One LLM key	See LLM providers — at least one is required

In production, also set:

Variable	Value	Why
`COOKIE_SECURE`	`true`	Forces session cookies to be HTTPS-only (recommended)

LLM providers

At least one provider key is required. Configuring a provider makes its models eligible for the picker — configure multiple to give agent authors a choice. Which models actually appear (and which one is preselected) is managed by workspace admins under Settings → Models; see Available models.

Provider	Environment variable	Models
Anthropic	`ANTHROPIC_API_KEY`	`claude-haiku-4-5`, `claude-sonnet-4-6`, `claude-sonnet-5`
OpenAI	`OPENAI_API_KEY`	`gpt-4o-mini`
Google	`GOOGLE_API_KEY`	`gemini-3-flash-preview`, `gemini-3-pro-preview`
DeepSeek	`DEEPSEEK_API_KEY`	`deepseek-v4-flash`, `deepseek-v4-pro`
Xiaomi	`XIAOMI_API_KEY`	`mimo-v2.5-pro`, `mimo-v2.5`
OpenRouter	`OPENROUTER_API_KEY`	`glm-5.2-max`, `glm-5.2-high`
Meta	`METAAI_API_KEY`	`muse-spark-1.2`

Anthropic models run with thinking enabled by default; Google models run with include_thoughts=True. Defaults live in app/model_providers/catalog.py.

Google via Application Default Credentials

GOOGLE_API_KEY is not the only way to enable Gemini. If it is unset but Application Default Credentials are available — a GCP service account on Cloud Run/GKE/Compute, gcloud auth application-default login, or GOOGLE_APPLICATION_CREDENTIALS — the backend serves Gemini models through Vertex AI using those credentials instead of the Gemini Developer API. The identity needs permission to call Vertex AI (e.g. the Vertex AI User role). When GOOGLE_API_KEY is set, it always wins.

Authentication

Variable	Notes
`GOOGLE_CLIENT_ID`	OAuth client ID for Google sign-in
`GOOGLE_CLIENT_SECRET`	OAuth client secret
`GOOGLE_REDIRECT_URI`	e.g. `https://your-host/auth/google/callback`
`AUTH_GOOGLE_EXCLUSIVE`	`true` disables email/password signup — Google OAuth only

Leaving these unset keeps email/password signup as the only path. See User Management for the full flow.

Sandbox (optional)

Agent code execution supports two providers, selected with:

Variable	Notes
`SANDBOX_PROVIDER`	`opensandbox` (default) or `cloudrun`

OpenSandbox

If you’re running OpenSandbox :

Variable	Notes
`OPEN_SANDBOX_DOMAIN`	e.g. `localhost:8083` or a private DNS name
`OPEN_SANDBOX_API_KEY`	If your OpenSandbox controller requires one
`OPEN_SANDBOX_USE_SERVER_PROXY`	`true` routes file ops through the OpenSandbox server (recommended)
`OPEN_SANDBOX_DEFAULT_IMAGE`	Base image, e.g. `python:3.12-slim`
`OPEN_SANDBOX_DEFAULT_PACKAGES`	Pre-installed packages, e.g. `["pandas","matplotlib"]`
`OPEN_SANDBOX_TIMEOUT`	Per-command timeout (seconds). Default `1800`
`OPEN_SANDBOX_VOLUME_MOUNTS`	Host mounts, comma-separated `host:sandbox[:ro]`

If OPEN_SANDBOX_DOMAIN is missing, the sandbox toggle on agents is silently disabled. Details: Sandbox.

Cloud Run sandboxes

If SANDBOX_PROVIDER=cloudrun, the backend drives Google Cloud Run sandboxes through the dedicated sandbox gateway service:

Variable	Notes
`CLOUD_RUN_SANDBOX_GATEWAY_URL`	URL of the deployed `sandbox-gateway` Cloud Run service
`CLOUD_RUN_SANDBOX_GATEWAY_SECRET`	Shared secret, sent as a bearer token — the gateway fails closed without it
`CLOUD_RUN_SANDBOX_GCS_BUCKET`	GCS bucket for turn-end snapshots (optional, enables cross-instance restore)
`CLOUD_RUN_SANDBOX_SNAPSHOT_PREFIX`	Object prefix inside the bucket. Default `sandbox-snapshots/`
`CLOUD_RUN_SANDBOX_ALLOW_EGRESS`	`true` requests outbound network per sandbox (gateway must also allow egress)
`CLOUD_RUN_SANDBOX_DEFAULT_PACKAGES`	Extra packages pip-installed into each fresh sandbox
`CLOUD_RUN_SANDBOX_TIMEOUT`	Per-command timeout (seconds). Default `1800`

Both CLOUD_RUN_SANDBOX_GATEWAY_URL and CLOUD_RUN_SANDBOX_GATEWAY_SECRET must be set for the sandbox toggle to be active. Details: Cloud Run Sandboxes.

Observability (optional)

Variable	Notes
`LANGFUSE_PUBLIC_KEY`	Langfuse project public key
`LANGFUSE_SECRET_KEY`	Langfuse project secret key
`LANGFUSE_BASE_URL`	`https://cloud.langfuse.com` or your self-hosted URL
`LANGFUSE_TIMEOUT`	API request timeout in seconds (positive integer; default `15`)

The public key, secret key, and base URL are required to enable Langfuse; missing any of them disables the integration silently. The timeout is optional. Details: Observability (Langfuse).

Slack (optional)

Variable	Notes
`SLACK_SIGNING_SECRET`	From Basic Information in your Slack app config
`SLACK_BOT_TOKEN`	`xoxb-…` Bot User OAuth Token

Missing values make the Slack endpoints reject all incoming requests. Details: Slack Integration.

Tuning & debugging

Variable	Default	Notes
`LOG_LEVEL`	`INFO`	Set to `DEBUG` for verbose logs (MCP traces, etc.)
`RECURSION_LIMIT`	`100`	Max LangGraph steps per agent invocation
`INVOKE_PROFILING`	`false`	Emit per-step timing on each agent invocation